The main objective of this project is to develop a statistical method of predicting assaultiveness in inpatient mentally-ill offenders. Assaultiveness is the result of a combination of variables from three domains: personal-trait, personal-state, and environmental/situational domains. We will construct a model predicting inpatient assaultiveness (yes/no and seriousness of assault) using specific variables from each of these domains and some interactions between variables across domains. Proposing to use hospital assaultiveness, within three months from hospital admission as the outcome variables, will allow the measurement of situational/environmental and patient-state factors surrounding the assault. The model will use patient-trait factors, measured on the evaluation battery, and patient-state and situational/environmental factors measured at times of assault. Some interactions between the patient-trait, patient-state, and environmental/situational factors will also be used. The variables of interest will be measured for both assaultive and non-assaultive study patients who share the same environment (ward). The model will be constructed using conditional logistic regression and multiple linear regression. The model will yield clinically useful information such as: given that a patient has trait factors A, B, and C, his relative risk of assault will change from X to Y given that patient-state and environmental factors change from a, b...,n to p,q...,t. Secondary goals are: a) To compare the statistical prediction with the clinical prediction made at the three-month mark; b) To test whether the in-hospital assaultive behavior is a valid predictor of community violent behavior. This will be done with incident reports for the patients who are stepped-down to civil hospitals and with RAP sheets for those who get discharged to the community. The subjects will consist of 400 consecutive consenting admissions to Kirby Forensic Psychiatric Center, a high security hospital for mentally-ill offenders in New York City. A measure of the intrinsic validity of the model will be obtained by using the "half and half" method.